simplifying surfaces. For a good overview of the topic the in-
terested reader is referred to [Heckbert and Garland, 1997].
In the Surface Adjacency Graph (SAG) paradigm, the points
are segmented into regions by fitting analytical surfaces to the
corresponding surface patches. An example of the application
of this paradigm is the region growing based on variable-order
surface fitting, such as the algorithm described in this paper.
2.2 Object recognition from laser scanning
Pre-processing, involving the computation of the laser foot-
print from the GPS, INS and laser range data, is usually fol-
lowed by thinning and manual editing to reduce the size of
the data stream and to remove the outliers. Then the cloud
of almost randomly distributed laser points is separated into
ground and non-ground returns by using low-level process-
ing, such as histogram thresholding, morphological filtering,
and the application of autoregressive integrated processes
([Lindenberger, 1993], and [Kraus and Pfeifer, 1998]). In-
stead of segmenting the whole surface into surface primitives,
the low-level processing is usually immediately followed by
model based object recognition and reconstruction of certain
type of objects, for example buildings (e.g., [Axelsson, 1999]
and [Maas and Vosselman, 1999]).
To facilitate the automation of the object recognition and
image understanding process many authors recommend the
inclusion of other sensory data into the process. The redun-
dant and complementary information from stereo photogram-
metry, multi(hyper) spectral imagery, and other sources can
be fused at the different points of the processing chain for
detection, and classification of features and objects, for sur-
face and objects reconstruction, and for error detection. For
example [Haala and Brenner, 1999] achieve automatic detec-
tion of.topographic features by combining laser data and color
imagery in a classification step, and uses laser data and 2D
ground plan information to obtain 3D reconstruction of build-
ings.
3 Segmentation and object recognition for different
applications
An ideal segmentation would partition the surface points into
surface primitives without making any domain-dependent as-
sumptions about specific objects, object classes, or applica-
tions [Besl, 1988]. In reality, the selection of the best seg-
mentation algorithm is application dependent. The sensor
characteristics, type of the measured quantities, the spatial
distribution of the laser points, the object properties, the er-
ror distribution, and the purpose of the survey all should be
considered.
Algorithms using surface fitting work well when recognition of
man-made objects, such as buildings, roads, etc., is the major
goal. These are smooth and solid objects usually bounded by
planar surfaces. The texture of surfaces (for example tiles on
a roof, or vegetation on topographic surfaces) are treated as
additional random noise superimposed on the smooth surface.
Segmentation of natural surfaces, such as sea ice, ice sheet or
land may pose different problems. These surfaces are viewed
as a combination of deterministic and stochastic parts. In
many applications, for example in geology and glaciology, the
statistical character of the surface (the stochastic part), is
also a subject of investigation. Surface texture statistics, like
surface roughness or correlation length can be the basis of
segmentation. Alternatively they can be used as additional
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
Internatioi
Figure 1: Aerial photograph of ice sheet surface with superglacial
lakes. The elevation profile in Figure 2 is extracted along the
white line. Solid arrows marks lake boundaries and the dashed
arrow points to ripples.
clues for object recognition.
Consider for example the area of superglacial lakes and com-
plicated drainage patterns on the Greenland ice sheet (Fig-
ure 1 and Figure 2(a). Although surface fitting is the logical
solution for describing the overall shape, ice sheets are so
simple that most of the object types can directly be recog-
nized by using simple clues, such as roughness and slope.
Figure 2 illustrates how scale-space analysis of the elevation
profiles as well as its first and second derivatives renders the
lakes. Lakes are horizontal and smooth, therefore the deriva-
tives of the surface elevation are small (for example see the
lake in Figure 1 and Figure 2, between the solid arrows).
The ripple zones, that is the series of bumps contouring the
lake shores down-glacier (for example around dashed arrow
in Figure 1), have large surface roughness resulting large and
rapidly alternating derivatives (Figure 2(b)-(c)). To suppress
this fine scale information the elevation profile was smoothed
by convolution with Gaussian kernels of increasing width. The
second derivatives of the smoothed elevation profiles was in-
spected to select the best scale for rendering the lake bound-
aries (Figure 2(d)).
A somewhat related topic is the determination of surface
statistics from a set of laser points or laser waveform. For hor-
izontal, random rough surfaces where the roughness scale is
much smaller than the footprint, the surface roughness within
each footprint can be estimated from the pulse width of the
return waveform [Gardner, 1992]. This approximation is valid
for some natural surfaces, like sea ice and for large footprint
laser system (LVIS, ICESAT). For more complicated surfaces
the approach introduced by [Goff and Jordan, 1988] might
be followed. He modeled the ocean floor as an anisotropic,
zero-mean, Gaussian random field to recover second order
statistics, namely amplitude, orientation, characteristic wave
numbers and Hausdorff (fractal) dimension of seafloor topog-
raphy from Sea Beam (sonar) data.
Figure 2: Object rec
on laser altimetry pro
mark lake boundaries
(a) Surface elevation
second derivatives, =
smoothed by convolu
4
The 0.1-02 m
service providers
[Krabill et al., 199!
low to medium fligl
reflectivity. In this
of a 0.05-0.15 m |
ranging (for GPS b
error of ca. 0.5-2
attitude errors and
[Schenk et al., 199
for horizontal and
sis demonstrates tl
topography, flight
and attitude deterr
No comprehensive
tifying the error of .
that larger errors te
ject boundaries sir
in larger elevation
faces. There is alsc
changes in elevatioi
footprint. This err:
ranging principle v
of the elevations v
noise. [Gardner, 1€
return waveforms t
ror on sloping or r:
broadening. His r
[Kraus and Pfeifer,
rain slope and roug
0.5-1 m for 1000 n
These results sugg:
